AI Briefing — March 8, 2026

🔬 Research

Heterogeneous Agent Collaborative Reinforcement Learning

141 upvotes · Zhixia Zhang, Zixuan Huang, Xin Xia, Deqing Wang, Fuzhen Zhuang · Paper

HACRL introduces a new paradigm where diverse RL agents collaborate during training by sharing verified rollouts, then operate independently at inference. Unlike traditional multi-agent RL requiring coordinated deployment, HACRL enables bidirectional learning among heterogeneous agents while using only half the rollout cost. The proposed HACPO algorithm outperforms GSPO by 3.3% on average across reasoning benchmarks, solving the sample efficiency problem in isolated on-policy optimization.

Helios: Real Real-Time Long Video Generation Model

133 upvotes · Shenghai Yuan, Yuanyang Yin, Zongjian Li, Xinwei Huang, Xiao Yang · Paper

The first 14B video generation model achieving 19.5 FPS on a single H100 GPU while supporting minute-scale generation without drift. Helios uses a unified autoregressive diffusion architecture for T2V, I2V, and V2V tasks, with training strategies that explicitly simulate drifting during training to prevent it at inference. The model fits four 14B instances in 80GB memory while matching quality of much larger baselines at computational costs comparable to 1.3B models.

T2S-Bench & Structure-of-Thought: Benchmarking and Prompting Comprehensive Text-to-Structure Reasoning

105 upvotes · Qinsi Wang, Hancheng Ye, Jinhee Kim, Jinghan Ke, Yifei Wang · Paper

Structure-of-Thought (SoT) prompting guides models to construct intermediate text structures, boosting performance across eight tasks and three model families. The accompanying T2S-Bench benchmark reveals substantial improvement potential with 1.8K samples across 6 scientific domains—even advanced models achieve only 58.1% node accuracy. On Qwen2.5-7B-Instruct, SoT alone provides +5.7% improvement, while fine-tuning on T2S-Bench increases this to +8.6%.

🛠 Tools

msitarzewski/agency-agents — Complete AI agency collection

1,468 stars/day · Repo

A curated collection of specialized AI agent personalities, each with unique expertise, communication styles, and production-ready workflows. Born from Reddit discussions and battle-tested, the 11K-star repository provides agents ranging from frontend developers to Reddit community managers, complete with deliverable-focused processes and success metrics. Designed to integrate with Claude Code and other AI platforms as reusable specialized characters rather than generic prompt templates.

openai/skills — Skills Catalog for Codex

948 stars/day · Repo

OpenAI's official skills catalog for Codex agents, providing folders of instructions, scripts, and resources that AI agents can discover and use for specific tasks. The 12.8K-star repository follows the Agent Skills open standard with system skills auto-installed in Codex, while curated and experimental skills can be installed via the built-in skill installer. Skills are designed to be written once and used everywhere, packaging repeatable capabilities for teams and individuals.

QwenLM/Qwen-Agent — Agent framework for Qwen models

586 stars/day · Repo

Agent framework built on Qwen>=3.0 featuring Function Calling, MCP (Model Context Protocol), Code Interpreter, RAG, and Chrome extension capabilities. The framework provides a comprehensive toolkit for building applications with Qwen models, targeting developers who need structured agent capabilities rather than basic chat interfaces. Reflects Alibaba's push to establish Qwen as a serious enterprise AI platform beyond just model releases.

📡 Analysis

Codex for Open Source

March 7, 2026 · Post · tags: open-source, ai, openai, generative-ai, llms, codex-cli

OpenAI launches six months of free ChatGPT Pro ($200/month) with Codex access for open source maintainers, following Anthropic's similar Claude Max offer from February. Unlike Anthropic's clear metrics (5K+ stars or 1M+ NPM downloads), OpenAI's application asks for "GitHub stars, monthly downloads, or why the project is important to the ecosystem." This represents escalating competition between AI providers for developer mindshare in the open source ecosystem.

Anthropic and the Pentagon

March 6, 2026 · Post · tags: bruce-schneier, ai, openai, generative-ai, llms, anthropic, ai-ethics

Bruce Schneier and Nathan E. Sanders analyze the Pentagon contracts with AI companies, highlighting how commoditized AI models make branding crucial for differentiation. As top-tier models from Anthropic, OpenAI, and Google achieve similar performance with minor leapfrogging improvements, Anthropic positions itself as the "moral and trustworthy" provider. This branding strategy has market value for both consumer and enterprise clients in an increasingly undifferentiated technical landscape.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net